Glossary term
Cluster (GPU/AI)
A networked group of GPUs working in parallel for training or inference; scale exposes bandwidth as the binding constraint.
A Cluster is a networked group of GPUs working in parallel on a single training or inference workload, ranging from a few hundred to over 100,000 accelerators. At scale, the interconnect bandwidth between GPUs becomes the binding constraint rather than the per-GPU compute. This is why frontier-model clusters drive disproportionate spending on InfiniBand, optical transceivers, and networking fabric. See Networking & Optical 101.